Repository navigation

All pull requests

Labels (56)Milestones (5)New pull request

[rl] Drain in-flight rollouts during shutdown
#5292felipemello1 opened Oct 11, 2026Contributor
[rl] Build and pack training batches in their own process
#5291felipemello1 opened Oct 11, 2026Contributor2 tasks
[rl] Controller GC: freeze the setup heap, fold generator metrics per rollout and group
#5290felipemello1 opened Oct 11, 2026Contributor
[rl] read each trainer rank's microbatches two ahead of training
#5286felipemello1 opened Oct 11, 2026Contributor
[rl] trainer ranks read their own microbatches by RDMA ref
#5285felipemello1 opened Oct 11, 2026Contributor3 tasks
11
Fix tuple assignment in spmd_types.py
#5283MaddipatlaChetan24 opened Oct 11, 2026Contributor
4
Initialize CPU-offloaded EMA directly in pinned host storage
#5282ZenAlexa opened Oct 11, 2026Contributor
6
[DistMuon] Send redistributed inputs and directions in BF16
#5279JavaZeroo opened Oct 11, 2026Contributor
3
[rl] TrainingSample per-token fields as tensors (no GC walk, torch.cat packing) (2/2)
#5274felipemello1 opened Oct 11, 2026Contributor
3
[rl] Store each turn's prompt as a delta on the previous turn (linear-size multi-turn rollouts) (1/2)
#5273felipemello1 opened Oct 11, 2026Contributor
3
[rl] RL model defaults in every run: fp32 lm_head, frozen MoE expert bias and router gate
#5271felipemello1 opened Oct 11, 2026Contributor
3
Make the vision encoder CP-invariant, only dp_shard
#5266acisseJZhong opened Oct 10, 2026Contributor
4
Fuse DSv3 MLA norm, MXFP8 and RoPE with CuTeDSL
#5265ydwu4 opened Oct 10, 2026Contributor
Fuse DSv3 HiMidLo router-gate backward experimentally
#5264ydwu4 opened Oct 10, 2026Contributor
[MoE] Opt-in shared-expert overlap with deferred DeepEP dispatch
#5263yuweih205 opened Oct 10, 2026
2
Fix dimension handling for batch invariant function
#5262MaddipatlaChetan24 opened Oct 10, 2026Contributor
3
Fuse DSv3 MTP cross-entropy forward/backward
#5261ydwu4 opened Oct 10, 2026Contributor
Add an opt-in DeepSeek V3 shared-expert autograd override
#5260ydwu4 opened Oct 10, 2026Contributor
Fuse DSv3 sequence-wise loss forward/backward
#5259ydwu4 opened Oct 10, 2026Contributor
1
Fuse DSv3 router forward/backward
#5258ydwu4 opened Oct 10, 2026Contributor
[rl] vLLM priority scheduling: serve rollout groups in the order they started
#5257felipemello1 opened Oct 10, 2026Contributor
[rl] keep session KV in an evict-last free-queue tier instead of pinning it (on #5220)
#5256pzhan9 opened Oct 10, 2026Contributor
[rl] MoE router replay: train on the experts the generator chose
#5255felipemello1 opened Oct 10, 2026Contributor5 tasks
7
[rl] Count every token in perf/trainer throughput; log controller GC time
#5254felipemello1 opened Oct 10, 2026Contributor
[rl] Move weight-sync draining from the router into the generator
#5253pzhan9 opened Oct 10, 2026Contributor